Vision detection method and system based on eye movement tracking trajectory analysis

Abstract

The application provides a vision detection method and system based on eye movement tracking trajectory analysis, which collects original eye movement trajectory data under a preset visual stimulus mode with a set size of an optotype and an observation distance; a plurality of intrinsic mode functions are obtained through variational mode decomposition, and a smooth following reference mode, a saccade mode of saccade movement and a noise residual mode are identified according to spectral and energy characteristics; a weighting matrix is constructed; a time-varying error function is constructed with the original data as input and the trajectory fitting error modulated by the weighting matrix as the target, and a zeroization neural network with a segmented non-monotonic activation function is used for iterative solution to obtain a smooth trajectory; the weighted deviation of the original fixation point and the smooth trajectory is calculated and compensated to obtain a corrected fixation point sequence; the deviation from the target position is calculated and converted into a converted visual angle error, and the vision detection result is determined in combination with the size of the optotype.

Description

Technical Field

[0001] This application belongs to the field of testing, and in particular relates to a vision testing method and system based on eye-tracking trajectory analysis. Background Technology

[0002] Eye-tracking technology enables objective, non-verbal visual acuity assessment by detecting the movement trajectory of the human eye when observing specific visual stimuli. However, in practical applications, limitations such as head micro-movements, eyelid occlusion, tear reflection, and the precision of the acquisition equipment result in raw eye-tracking trajectory data containing not only smooth following and saccadic movements representing visual intent but also random noise. This noise intertwines with the actual physiological characteristics of eye movements, leading to distortion in fixation point calculations if visual acuity is determined based on the raw eye-tracking trajectory, thus reducing the accuracy and clinical value of objective visual acuity testing. This paper combines signal decomposition techniques with neural network models, using modal decomposition algorithms to break down non-stationary eye-tracking signals into components of different frequency bands. Neural networks, adept at handling time-varying signals, are then used to fit and optimize the trajectory, reconstructing a smooth and realistic fixation path. However, existing methods lack weighting mechanisms after decomposing the eye-tracking signal, failing to distinguish and extract smooth following, saccadic, and noise components based on the signal's spectral and energy characteristics, and thus unable to utilize intermodal correlations to construct reasonable confidence levels. When using a neural network to solve for time-varying trajectory errors, the conventional monotonic activation function results in slow network convergence and insufficient convergence accuracy, failing to bring the time-varying error to zero in real time. After obtaining the optimized trajectory, the spatial weighted deviation of the fixation point was not correlated and compensated with the set key parameters of observation distance and optotype size. Consequently, the calculated trajectory deviation could not be converted into a visual angle error that meets clinical standards, and reliable visual acuity test results could not be output. Summary of the Invention

[0003] This invention provides a vision detection method and system based on eye-tracking trajectory analysis.

[0004] According to one aspect of the present invention, a vision detection method based on eye-tracking trajectory analysis is provided, the method comprising:

[0005] Acquire raw eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance; perform variational mode decomposition on the raw eye movement trajectory data to obtain multiple intrinsic mode functions, and identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral and energy characteristics.

[0006] A noise suppression factor is determined based on the residual modal components, a fidelity factor is determined based on the baseline modal components, and a saccade compensation factor is determined based on the saccade modal components. The product of the noise suppression factor, the fidelity factor, and the saccade compensation factor is used as the confidence level to construct a weighted matrix. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the goal of minimizing the trajectory fitting error modulated by the weighted matrix. A nulled neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory.

[0007] The weighted deviation between the fixation point and the corresponding position of the smooth trajectory in the original eye movement trajectory data is calculated, and a corrected fixation point sequence is obtained after compensation. The degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern is calculated, and the degree of deviation is converted into a calculated visual angle error in combination with the set observation distance. The visual acuity test result is determined by combining the calculated visual angle error and the set optotype size.

[0008] According to another aspect of the present invention, a vision detection system based on eye-tracking trajectory analysis is provided, the system comprising the following modules:

[0009] The recognition module is used to acquire the original eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance; to perform variational mode decomposition on the original eye movement trajectory data to obtain multiple intrinsic mode functions, and to identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral characteristics and energy characteristics.

[0010] The solution module is used to determine a noise suppression factor based on the residual modal components, a fidelity factor based on the reference modal components, and a saccade compensation factor based on the saccade modal components. The product of the noise suppression factor, the fidelity factor, and the saccade compensation factor is used as a confidence level to construct a weighted matrix. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the objective of minimizing the trajectory fitting error modulated by the weighted matrix. A nullable neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory.

[0011] The determination module is used to calculate the weighted deviation between the fixation point in the original eye movement trajectory data and the corresponding position of the smooth trajectory, and obtain a corrected fixation point sequence after compensation; calculate the degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern, convert the degree of deviation into a calculated visual angle error in combination with the set observation distance, and determine the visual acuity test result by combining the calculated visual angle error and the set optotype size.

[0012] This invention performs variational mode decomposition on raw eye movement trajectory data, separating the baseline, saccade, and residual mode components by combining spectral and energy features, thus removing interference noise and preserving authentic eye movement characteristics. A comprehensive confidence and weighting matrix is ​​constructed by combining noise suppression and fidelity factors, ensuring high signal fidelity while suppressing noise. A time-varying error function modulated by this weighting matrix is ​​used, and a nullified neural network with piecewise non-monotonic activation functions is employed for iterative solution to obtain a highly smooth eye movement trajectory. By scientifically compensating for fixation point deviation and converting the deviation into a calculated visual angle error, visual acuity test results can be obtained without subjective feedback from the test subject, improving the objectivity and anti-interference capability of visual acuity assessment. Attached Figure Description

[0013] Figure 1 This is a flowchart of a vision testing method based on eye-tracking trajectory analysis;

[0014] Figure 2 This is a schematic diagram of the time-domain waveforms of each modal component after variational mode decomposition of the original eye-tracking trajectory data;

[0015] Figure 3 This is a schematic diagram of the iterative convergence curve of a nullified neural network with piecewise non-monotonic activation functions.

[0016] Figure 4 This is a bar chart comparing the core performance indicators of the proposed solution with those of traditional algorithms. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0018] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0019] Example 1

[0020] In Embodiment 1 of the present invention, as Figure 1 As shown, a vision detection method based on eye-tracking trajectory analysis includes:

[0021] S1, acquire raw eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance.

[0022] Preset visual stimulus patterns are presented on a computer screen. A psychological toolkit is used to generate Landau ring optotypes with different spatial frequencies, and the distance from the center of the monitor to the subject's eyes is set as the target viewing distance. Visual stimulus patterns include, but are not limited to, a smooth tracking mode, where a dot appears on the screen moving at a constant speed or a sinusoidal pattern, and the subject's gaze follows it continuously; in a vision testing mode, an E-shaped optotype with a specific opening direction appears at the center of the screen or a designated location at a set viewing distance and size; in a reverse saccade, a bright dot suddenly flashes on one side of the screen, and the subject's gaze quickly jumps to the opposite side. The two-dimensional screen coordinates of the subjects' eyes while viewing these patterns, along with the corresponding timestamps, are collected as raw eye movement trajectory data and stored in a file format using a generic data processing toolkit.

[0023] S2, perform variational mode decomposition on the original eye-tracking trajectory data to obtain multiple intrinsic mode functions, and identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral and energy characteristics.

[0024] Using the variational mode decomposition function of the vmdpy library, the x and y axes of the original eye-tracking data were decomposed separately. A penalty factor of 2000 and a decomposition level of 5 were set to obtain five intrinsic mode functions (IMFs). The power spectral density of each IMF was calculated, the center frequency was extracted as the spectral feature, and the root mean square value of each mode signal was calculated as the energy feature. Modes with a center frequency below 5Hz and whose energy features accounted for more than 80% of the total energy were identified as baseline mode components. Modes with a center frequency between 5Hz and 20Hz and exhibiting abrupt energy peaks were identified as saccade mode components. Modes with a center frequency above 20Hz and whose energy features accounted for less than 5% of the total energy were identified as residual mode components. Figure 2 As shown in the figure, the time-domain waveforms of the original eye-tracking trajectory data, the reference modal component corresponding to the smooth following motion, the saccade modal component corresponding to the saccade motion, and the residual modal component corresponding to the noise are displayed within a 2-second time period.

[0025] In some embodiments, performing variational mode decomposition on the original eye-tracking trajectory data to obtain multiple intrinsic mode functions includes:

[0026] Set the penalty factor for variational mode decomposition and the number of intrinsic mode functions to be extracted;

[0027] The decomposition process is constructed as a constrained variational extremum solution process, and the alternating direction multiplier method is used to update the center frequency and signal bandwidth of each decomposition state in the frequency domain one by one.

[0028] Based on the updated center frequency, time-domain sequence reconstruction is performed in each frequency band, and isolated sequence data is stripped away to output the intrinsic mode functions corresponding to multiple channels without cross-aliasing characteristics.

[0029] When performing variational mode decomposition (VMD) on the raw eye movement trajectory data, a VMD penalty factor is set based on the complexity and signal-to-noise ratio of the eye movement signal. A preferred range is 1500 to 2500, with an example value of 2000. This parameter balances bandwidth fidelity and noise tolerance during the decomposition process. Simultaneously, the number of intrinsic mode functions (IMFs) to be extracted, K, is set, typically 3 to 5 based on the main characteristics of eye movement; an example setting is 4. A tolerance parameter, for example, 0.001, is initialized.

[0030] By utilizing the augmented Lagrangian function, the decomposition process is constructed as a constrained variational extremum solution process that minimizes the sum of the bandwidths of each mode. In the iterative solution phase, the alternating direction multiplier method is used to alternately update the eigenmode functions, center frequencies, and Lagrangian multipliers of each decomposition state in the frequency domain. For each mode, the amplitude expression in the frequency domain is updated using the Wiener filter computational architecture, and the center frequency and corresponding signal bandwidth are calculated and updated in real time according to the energy centroid method formula. When the rate of change of the center frequency of each mode in two consecutive iterations is less than the set convergence tolerance... The iteration terminates when the time is right.

[0031] Based on the center frequencies obtained through iterative convergence, time-domain sequence reconstruction is performed in each frequency band of the frequency domain through inverse fast Fourier transform. The background baseline component is removed from the time-domain reconstructed sequence to isolate the sequence data. K intrinsic mode functions with narrow bandwidth characteristics and no cross-aliasing are stably output for multiple channels, thus transforming the original non-stationary eye-tracking signal into a stationary frequency band sequence.

[0032] In some embodiments, identifying the reference mode component of the smooth following motion, the saccade mode component of the saccade motion, and the residual mode component of the noise based on spectral and energy characteristics includes:

[0033] Calculate the signal dominant frequency of each intrinsic mode function and the proportion of each intrinsic mode function in the overall energy distribution in a single extraction segment;

[0034] The intrinsic mode functions with a dominant frequency of less than 2 Hz and an energy percentage of more than 50% are used as the reference mode components;

[0035] The intrinsic mode function containing extreme velocity abrupt changes and exhibiting a broadband spectrum distribution is used as the scanned mode component;

[0036] The intrinsic mode functions with a dominant frequency greater than 15Hz and frequency energy divergence exceeding a preset threshold are used as the residual mode components.

[0037] The Fast Fourier Transform (FFT) algorithm is used to transform the intrinsic mode functions (EMFs) in the time domain to the frequency domain. The dominant frequency of the signal is calculated by finding the frequency point corresponding to the maximum spectral amplitude. Simultaneously, the sum of squares of the EMFs within a single extraction time window (e.g., a 2-second window) is calculated as a sequence integral. This integral is then divided by the sum of all mode integrals to determine the percentage of each EMF in the overall energy distribution.

[0038] During feature recognition, a filtering logic is set to extract intrinsic mode functions with a main frequency in the preferred range of 0.5 to 1.5 Hz (e.g., 1.2 Hz) and a calculated energy percentage greater than 50% (e.g., 68% in the example). These are used as the reference mode components to represent the smooth following motion of the eye following the target at low speed.

[0039] Simultaneously, the instantaneous eye movement velocity is extracted by performing first-order time difference operation on the intrinsic mode function, and the abrupt change point in the waveform that exceeds the preset peak velocity threshold, such as a step extreme value of 300° per second, is identified. The power spectrum is observed to cover multiple sub-bands from 2 to 20 Hz, thus exhibiting a broadband distribution characteristic. The intrinsic mode function that satisfies these two conditions is confirmed as the saccadic mode component representing rapid gaze skipping motion.

[0040] For the main frequency being in the high-frequency range, i.e. greater than 15Hz, preferably 20 to 30Hz, with an example value of 25Hz, and the frequency energy divergence calculated using the frequency band variance algorithm exceeding a preset threshold, the example threshold value is set to an intrinsic mode function of 0.1 watts per Hz, which is used as a residual mode component that is mixed with eye tremor and environmental power frequency electromagnetic interference for stripping.

[0041] S3. Determine the noise suppression factor based on the residual modal components, determine the fidelity factor based on the reference modal components, and determine the saccade compensation factor based on the saccade modal components. Construct a weighted matrix by multiplying the noise suppression factor, the fidelity factor, and the saccade compensation factor as the confidence level.

[0042] A sliding window algorithm is used to calculate the local variance of the residual modal components at each sampling time. The reciprocal of this local variance is then smoothed and mapped to the 0-1 interval using a natural exponential function to obtain a noise suppression factor. Simultaneously, the Pearson correlation coefficient function from the numerical computation library is used to calculate the absolute value of the linear correlation between the baseline modal components and the original eye-tracking trajectory data within the corresponding sliding window, serving as a fidelity factor. Simultaneously, the instantaneous energy amplitude of the saccade modal components within the corresponding sliding window is calculated and converted into a saccade compensation factor using an inverse proportional attenuation mapping. At each time step, the corresponding noise suppression factor, fidelity factor, and saccade compensation factor are multiplied to calculate the confidence level for each sampling point. A diagonal matrix generation function from the numerical computation library is used to construct a diagonal weighted matrix with the same dimension as the data length, using the confidence levels of all time steps as the main diagonal elements.

[0043] In some embodiments, determining a noise suppression factor based on the residual mode components, determining a fidelity factor based on the reference mode components, and determining a saccade compensation factor based on the saccade mode components, and constructing a weighted matrix using the product of the noise suppression factor, the fidelity factor, and the saccade compensation factor as confidence levels, includes:

[0044] The ratio of the absolute value of the residual modal component to the absolute value of the original eye-tracking trajectory data is used as the base ratio. The noise suppression factor is obtained by exponentiation with the natural constant as the base and the negative number of the base ratio as the exponent.

[0045] Calculate the ratio of the time derivative of the reference mode component at the current moment to the preset reference velocity, perform a hyperbolic tangent transformation on the absolute value of the ratio, and subtract the transformation result from 1 to obtain the fidelity factor;

[0046] Calculate the ratio of the absolute value of the time derivative of the saccade modal component at the current moment to the preset saccade threshold, and map the ratio to the saccade compensation factor using a Gaussian decay function;

[0047] Multiply the noise suppression factor, the fidelity factor, and the saccade compensation factor at the current moment to obtain the confidence level at the current moment;

[0048] Construct a diagonal matrix using the confidence scores at each time point as diagonal elements, and use the diagonal matrix as the weighting matrix.

[0049] In the calculation of confidence weights, the instantaneous absolute amplitude of the residual modal component is calculated in each sampling period and divided by the absolute amplitude of the original eye-tracking trajectory data at the same moment. The resulting value is used as the basic proportion variable representing the relative intensity of local high-frequency interference. The value of this variable usually fluctuates between 0 and 0.4.

[0050] The exponential decay output is limited to a noise suppression factor within the range greater than 0 and less than or equal to 1, causing the suppression coefficient at high-noise moments to decay close to zero. When constructing the tracking fidelity mechanism, the time derivative of the reference modal component at the current moment is obtained using an interpolation method with adjacent sampling points at a step size of 10ms. This instantaneous tracking velocity is then divided by a preset reference velocity representing the visual tracking baseline, preferably set within a range of 5 to 15 degrees per second (e.g., 10 degrees per second), to obtain the ratio. A tanh hyperbolic tangent transform is performed on the absolute value of this ratio to smoothly compress drastic fluctuations into a range greater than or equal to 0 and less than 1. The fidelity factor is obtained by subtracting 1 from the mapping result. This ensures that when the eye is in a stable tracking or fixation state, the hyperbolic tangent transform value approaches 0, and after subtraction, the fidelity factor approaches 1, thus giving the highest confidence to stable eye-tracking data. Conversely, when the eye-tracking velocity undergoes abnormal abrupt changes, the fidelity factor automatically decays and approaches 0, thereby suppressing low-fidelity abnormal tracking states.

[0051] To dynamically adjust the weights during saccades, the absolute value of the time derivative of the saccade modal component at the current moment is calculated as the saccade change rate. This rate is divided by a preset saccade threshold, for example, 500° per second, to obtain a ratio. This ratio is then mapped using a Gaussian decay function with a natural constant as the base, compressing it to a range greater than 0 and less than or equal to 1 to obtain the saccade compensation factor. This ensures that the confidence level decreases smoothly with increasing saccade intensity during rapid saccades, preventing excessive distortion of the trajectory fitting at points of abrupt displacement changes.

[0052] Perform element-wise dot product operations on vectors, multiplying the noise suppression factor, fidelity factor, and saccade compensation factor calculated at the current time step, and outputting an instantaneous comprehensive confidence value of, for example, 0.65. Arrange the confidence scores of all N discrete time steps extracted within the sampling segment sequentially on the main diagonal of an N×N dimension matrix, and clear and fill the remaining off-diagonal matrix elements to generate the weighted matrix used to modulate the residuals of the solution model.

[0053] S4. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the goal of minimizing the trajectory fitting error modulated by the weighting matrix; a nulled neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory.

[0054] A time-varying error function is constructed by defining the difference vector between the unknown smooth trajectory variable and the original eye-tracking trajectory data and left-multiplying it by the aforementioned diagonal weighting matrix. Based on the principle of nullable neural networks, the time derivative of this time-varying error function is set to the negative convergence coefficient multiplied by the activation function output value, thus establishing a differential equation variation model for the neural network. The activation function uses a sine function for smoothing in the interval where the absolute error value is less than 1, and a sign function with an exponentially decaying term is used in the interval where the absolute error value is greater than or equal to 1 to achieve non-monotonic adjustment. The fourth-order, fifth-level Runge-Kutta ordinary differential equation solver in the scientific computing library's integration module is used to numerically integrate and solve the neural network variation model. The trajectory vector output after system state convergence is taken as the smoothed trajectory.

[0055] In some embodiments, the step of using a nullable neural network with piecewise non-monotonic activation functions to iteratively solve the time-varying error function to obtain a smooth trajectory includes:

[0056] A neural dynamics equation is constructed based on the principle of null neural networks, and the time-varying error function and time derivative are substituted into the neural dynamics equation.

[0057] Set a convergence threshold and a segmented interval for the activation function. During the iterative solution process, calculate the time-varying error function value at the current moment. Adjust the mapping parameters of the segmented non-monotonic activation function according to the segmented interval where the time-varying error function value is located.

[0058] The time-varying error function value is input into the piecewise non-monotonic activation function after parameter adjustment, and the state variables of the null neural network are updated.

[0059] The process continues iteratively until the time-varying error function value is less than the convergence threshold. The output state variable of the nullified neural network at this point is then used as the smoothed eye-tracking position coordinates, and the smoothed trajectory is formed based on the time series.

[0060] Based on the control theory principle that forces the error matrix of a nullable neural network to asymptotically approach zero, a neurodynamic model equation is established in the continuous time domain. The time derivative of the error is equal to the error itself under the control of the negative feedback activation function. The nullable neural network has a single-layer feedforward and integral feedback structure, and the network structure includes a signal input layer, an execution computation hidden layer, and a state output layer. The input of the network model is the time-varying error function of the eye movement trajectory modulated by a weighted matrix and the time derivative obtained by numerical differentiation using the backward finite difference method. The execution computation hidden layer incorporates the piecewise non-monotonic activation function to perform forward mapping calculations, implements gain control of the error signal through the neurodynamic equation, and outputs feedback driving variables. The state output layer contains a time integration unit that accumulates driving variables and continuously updates the internal state. The output of this network model is the smoothed two-dimensional eye movement position coordinates after eliminating jitter noise. Based on the above network structure, the constructed neurodynamic equation is expressed as follows: Where E(t) represents the time-varying error function as an input variable. The time derivative of the time-varying error function is also used as an input variable. Indicates the internal gain mapping parameters. This represents the piecewise non-monotonic activation function.

[0061] A pre-written memory constant is used as the convergence threshold, preferably set within a certain range. to Between, for example, take As a hard cutoff parameter, the error range is divided into intervals. The segmented intervals of the activation function are divided into a large error region (error absolute value is greater than 1), a medium error region (error absolute value is greater than 0.1 and less than or equal to 1), and a small error region (error absolute value is less than or equal to 0.1).

[0062] In the inner loop of the iterative solution using the fourth-order Runge-Kutta numerical method, the time-varying error function at the current time step is calculated at each integration step. Absolute norm value. Based on the specific segmented interval into which the function value falls, look up a table and adjust the internal gain mapping parameter of the segmented non-monotonic activation function. When it is determined to be in the large error region, assign a higher exponential convergence coefficient, such as 3.0. When it shrinks to the small error region, smoothly switch to a non-monotonic linear coefficient that incorporates sinusoidal perturbations, such as 1.0, to prevent local oscillations from causing the function to get stuck.

[0063] The current error value is fed into the piecewise non-monotonic activation function with the latest parameters for forward calculation, generating driving variables to fine-tune the internal state variables in the null neural network memory in real time. This stepping process continuously iterates and updates at high speed according to a simulation step size of 0.01s, constantly monitoring the error indicators until the time-varying error function value is compressed and less than the set threshold. When the convergence threshold is reached, a stop signal is triggered at that point. The output state variable of the null neural network that has reached convergence is the smoothed two-dimensional eye-tracking position coordinate. The coordinates of each point throughout the entire time period are reconstructed and assembled into the complete smoothed trajectory based on the original timestamp sequence. Figure 3 As shown in the figure, the time-varying error function changes with the number of iterations, and it can be seen that the network error approaches 0 after about 40 iterations.

[0064] S5, calculate the weighted deviation between the gaze point and the corresponding position of the smooth trajectory in the original eye movement trajectory data, and obtain the corrected gaze point sequence after compensation.

[0065] A velocity-threshold-based gaze point recognition algorithm is used to classify the raw eye movement trajectory data, extracting discrete gaze points with velocities below 30° of visual angle per second. The Euclidean distance function from a numerical computation library is used to calculate the two-dimensional spatial difference between the coordinates of these discrete gaze points and the corresponding coordinates of the smooth trajectory at the same timestamp. This difference is then multiplied by the corresponding confidence level in a diagonal weighted matrix to obtain a weighted bias. This weighted bias is then inversely superimposed onto the original discrete gaze point coordinates for coordinate translation compensation, resulting in a corrected gaze point sequence that eliminates high-frequency jitter and device system drift.

[0066] In some embodiments, calculating the weighted deviation between the fixation point and the corresponding position of the smoothed trajectory in the original eye movement trajectory data, and obtaining the corrected fixation point sequence after compensation, includes:

[0067] Obtain the original gaze point coordinates at the current moment in the original eye movement trajectory data, and the smooth point coordinates at the current moment in the smooth trajectory;

[0068] Calculate the vector difference between the original gaze point coordinates and the smoothed point coordinates to obtain the position deviation vector;

[0069] Calculate the ratio of the magnitude of the position deviation vector to the preset reference attenuation radius, calculate the attenuation value of the ratio using the exponential attenuation function, subtract the attenuation value from 1 to obtain the deviation weight, and multiply the position deviation vector by the deviation weight to obtain the weighted deviation.

[0070] Subtract the weighted deviation from the original fixation point coordinates to obtain the corrected fixation point coordinates at the current time. Sort the corrected fixation point coordinates at each time according to time to form the corrected fixation point sequence.

[0071] In the spatial domain compensation operation of coordinate error, the synchronous matching module obtains the unfiltered two-dimensional original gaze point coordinate data of the current recording time in the original eye movement trajectory data in parallel according to the global timestamp, and synchronously extracts the smoothed point coordinate data belonging to the same time section in the smoothed trajectory from the network solution cache pool.

[0072] The computational core calculates the algebraic difference between the x-axis and y-axis coordinate components, that is, subtracting the x-axis and y-axis values ​​of the corresponding smoothed point from the original gaze point coordinates to obtain a two-dimensional positional deviation vector representing the direction and magnitude of the absolute displacement. A preset reference attenuation radius constant representing the foveal tolerance of the retina is loaded, with the preferred parameter range set to 0.5 to 2.0 spatial diopters; 1.0° is used in the example. The Euclidean modulus of the extracted positional deviation vector is calculated using the Pythagorean theorem and divided by the reference attenuation radius to obtain a dimensionless spatial normalization ratio.

[0073] Based on this, the natural decay value of the ratio is calculated using the standard negative exponential decay formula, and the deviation weighting coefficient, which is limited to a surface greater than or equal to 0 and less than 1, is calculated by subtracting this decay value from the integer 1.

[0074] By multiplying the positional deviation vector by the obtained deviation weight using a vector scalar multiplication operation, a weighted deviation for compensation based on displacement amplitude scaling is obtained. Using basic vector subtraction, this weighted deviation variable is subtracted from the initially acquired original fixation point coordinates, thereby eliminating coordinate tremor errors caused by blinking or electrophysiological noise, generating corrected fixation point coordinates at the sampling point, and performing cyclic compensation throughout the entire detection period. The corrected fixation point sequence is then reassembled sequentially to obtain the sequence specifically for visual function evaluation.

[0075] S6, calculate the degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern, convert the degree of deviation into a calculated visual angle error in combination with the set observation distance, and determine the visual acuity test result by combining the calculated visual angle error and the set optotype size.

[0076] The true screen coordinates of the Landau ring target notch center in the preset visual stimulus pattern are extracted. The two-dimensional centroid position of the corrected fixation point sequence is calculated using a numerical calculation library, and the screen pixel Euclidean distance between this centroid position and the true screen coordinates is obtained. Based on the screen resolution and size, the pixel distance is converted to millimeter distance. The arctangent function from the mathematical library is used to calculate the arctangent value of the ratio of this millimeter distance to the set observation distance, yielding the converted visual angle error in degrees. This converted visual angle error is input into a pre-trained multiple linear regression model using the least squares method. This model uses the converted visual angle error and the spatial frequency corresponding to the set target size as variables to calculate and output the subject's logarithmic minimum visual acuity value as the visual acuity test result.

[0077] In some embodiments, calculating the deviation of the corrected fixation point sequence from the target position in the preset visual stimulus pattern, converting the deviation into a calculated visual angle error in conjunction with the set observation distance, and determining the visual acuity test result by combining the calculated visual angle error with the set optotype size includes:

[0078] Extract the true coordinate sequence of the preset target point at each time moment in the preset visual stimulus pattern;

[0079] Calculate the Euclidean distance between each fixation point in the corrected fixation point sequence and the true coordinates at the corresponding time, and use it as the instantaneous deviation error at each time.

[0080] The average deviation error is calculated by discretely summing the instantaneous deviation errors at all times and dividing by the total number of sampling points.

[0081] The average deviation error is converted into a calculated visual angle error by combining the set observation distance. The calculated visual angle error is compared with the preset stable fixation angle threshold. Based on the set target size and set observation distance corresponding to the preset visual stimulation mode when the stable fixation condition is reached, the corresponding visual acuity test level is determined and output as the visual acuity test result.

[0082] When entering the objective evaluation and conversion process for clinical visual acuity parameters, the assessment software synchronously extracts the sequence of true coordinates of preset target points corresponding to each absolute time point on the screen display matrix of the preset visual stimulus pattern presented during the subject's observation phase from the feedback log of the visual stimulus rendering engine. For example, the absolute pixel position of the target centroid. For the time-aligned slices throughout the test, the absolute Euclidean distance between the two-dimensional coordinates of the viewpoint extracted at each moment in the corrected fixation point sequence and the corresponding true coordinates of the target is calculated using the distance formula between two points. The resulting scalar length unit is screen measurement pixels or millimeters. For example, a current deviation of 4.2mm is recorded as the single-frame instantaneous deviation error for this test individual.

[0083] The instantaneous deviation errors at all times corresponding to the current evaluation segment, such as 1500 consecutive sampling points, are fed into a discrete accumulator for summation. The resulting scalar sum is divided by the number of sampling points to obtain the average deviation error value. The set observation distance variable, representing the straight-line span from the cornea to the stimulation plane, is retrieved from the rangefinder. The preferred distance is specified as being between 400mm and 600mm, for example, using 500mm as the parameter input in routine ophthalmological examinations. The previously obtained average deviation error is substituted into the geometric arctangent function to perform the calculation and conversion formula. ,in Let e ​​be the visual angle, d be the average error, and d be the observation distance. The linear deviation is converted into a visual angle error in radians or degrees, unconstrained by the size of the display device; for example, a parallax of 0.35° is calculated. The calculated visual angle error is input into the decision tree node and compared with a preset stable fixation angle threshold constant used in medicine to determine visual fixation status, typically set between 0.5° and 1.5°. An example threshold configuration of 0.6° is used for comparison verification. In one embodiment, different test users correspond to different preset stable fixation angle thresholds. If the calculated error is less than or equal to this threshold, a stability judgment logic is triggered. Then, the specific optotype parameters given when the preset visual stimulation mode is triggered in this round are obtained, such as the Landau ring feature with a notch size of 1.45mm and the set observation distance at that time. These parameters are substituted into a standard ophthalmological conversion formula to deduce the subject's resolution limit, determining the corresponding standardized grade score, such as visual acuity level 5.0 or decimal 1.0. This digital clinical grade is written into the health report as the verified visual acuity test result.

[0084] The experiment was conducted in a controlled optical darkroom, with an observation distance of 500mm, a stimulation screen refresh rate of 120Hz, and an eye tracker sampling frequency of 1000Hz. One hundred subjects without significant organic ocular lesions were recruited. Power frequency electromagnetic interference was used, and subjects were guided to generate micro-tremors in their heads. The experiment was divided into a control group and an experimental group based on the proposed protocol. The control group used traditional bandpass filtering combined with Kalman smoothing to process the raw eye-tracking data. The experimental group underwent a complete processing procedure including variational mode decomposition, construction of a weighted matrix confidence level, iterative solution of the smoothed trajectory using a null neural network, and weighted bias spatial compensation. Comparison was based on the gold standard visual acuity test results issued by a clinically qualified physician. Figure 4 As shown in the figure, the test results of the control group and the experimental group are compared on three core indicators: offset error, angle error, and visual acuity detection accuracy.

[0085] After collecting 50 hours of eye-tracking test data, the statistical analysis phase commenced. Regarding specific data and experimental results, the average deviation error calculated for the experimental group in this application was 2.15 mm, while that for the control group was 4.82 mm. After geometric arctangent transformation, the calculated visual angle error for the experimental group stabilized at 0.24°, while the control group's error reached as high as 0.55°, and the control group exhibited failure judgments exceeding the 0.6° stable fixation angle threshold. In terms of visual acuity testing accuracy, the experimental group's matching degree with the medical gold standard reached 96.5%, while the control group's visual acuity testing accuracy was only 81.2%. Furthermore, when faced with rapid saccadic movements and micro-flickering at 300° per second, the root mean square error of eye trajectory reconstruction in the experimental group decreased to 0.013°.

[0086] Traditional algorithms in the control group cannot separate broadband eye-tracking signals from high-frequency electromagnetic noise, leading to eye tremor interference that disrupts the smoothness of the temporal trajectory and reduces the accuracy of visual acuity assessment. This application's solution utilizes variational mode decomposition to extract baseline and saccade mode components, and removes residual mode components. Based on attenuation algorithms, integration mechanisms, and saccade compensation to generate confidence scores, it successfully achieves noise-resistant weighting and dynamic adaptation against local high-frequency interference and drastic saccade abrupt changes. Simultaneously, the null neural network combined with piecewise non-monotonic activation functions enables millisecond-level switching of mapping gain parameters when facing different error conditions (large and small errors). This eliminates coordinate tremors in the original data and suppresses local oscillation deadlocks in the algorithm's internal loop, achieving a reduction in visual angle error and improving the accuracy of clinical optotype assessment.

[0087] Example 2

[0088] Embodiment 2 of the present invention proposes a vision detection system based on eye-tracking trajectory analysis, comprising the following modules:

[0089] The recognition module is used to acquire the original eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance; to perform variational mode decomposition on the original eye movement trajectory data to obtain multiple intrinsic mode functions, and to identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral characteristics and energy characteristics.

[0090] The solution module is used to determine a noise suppression factor based on the residual modal components, a fidelity factor based on the reference modal components, and a saccade compensation factor based on the saccade modal components. The product of the noise suppression factor, the fidelity factor, and the saccade compensation factor is used as a confidence level to construct a weighted matrix. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the objective of minimizing the trajectory fitting error modulated by the weighted matrix. A nullable neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory.

[0091] The determination module is used to calculate the weighted deviation between the fixation point in the original eye movement trajectory data and the corresponding position of the smooth trajectory, and obtain a corrected fixation point sequence after compensation; calculate the degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern, convert the degree of deviation into a calculated visual angle error in combination with the set observation distance, and determine the visual acuity test result by combining the calculated visual angle error and the set optotype size.

[0092] In some embodiments, performing variational mode decomposition on the original eye-tracking trajectory data to obtain multiple intrinsic mode functions includes:

[0093] Set the penalty factor for variational mode decomposition and the number of intrinsic mode functions to be extracted;

[0094] The decomposition process is constructed as a constrained variational extremum solution process, and the alternating direction multiplier method is used to update the center frequency and signal bandwidth of each decomposition state in the frequency domain one by one.

[0095] Based on the updated center frequency, time-domain sequence reconstruction is performed in each frequency band, and isolated sequence data is stripped away to output the intrinsic mode functions corresponding to multiple channels without cross-aliasing characteristics.

[0096] In some embodiments, identifying the reference mode component of the smooth following motion, the saccade mode component of the saccade motion, and the residual mode component of the noise based on spectral and energy characteristics includes:

[0097] Calculate the signal dominant frequency of each intrinsic mode function and the proportion of each intrinsic mode function in the overall energy distribution in a single extraction segment;

[0098] The intrinsic mode functions with a dominant frequency of less than 2 Hz and an energy percentage of more than 50% are used as the reference mode components;

[0099] The intrinsic mode function containing extreme velocity abrupt changes and exhibiting a broadband spectrum distribution is used as the scanned mode component;

[0100] The intrinsic mode functions with a dominant frequency greater than 15Hz and frequency energy divergence exceeding a preset threshold are used as the residual mode components.

[0101] In some embodiments, determining a noise suppression factor based on the residual mode components, determining a fidelity factor based on the reference mode components, and determining a saccade compensation factor based on the saccade mode components, and constructing a weighted matrix using the product of the noise suppression factor, the fidelity factor, and the saccade compensation factor as confidence levels, includes:

[0102] The ratio of the absolute value of the residual modal component to the absolute value of the original eye-tracking trajectory data is used as the base ratio. The noise suppression factor is obtained by exponentiation with the natural constant as the base and the negative number of the base ratio as the exponent.

[0103] Calculate the ratio of the time derivative of the reference mode component at the current moment to the preset reference velocity, perform a hyperbolic tangent transformation on the absolute value of the ratio, and subtract the transformation result from 1 to obtain the fidelity factor;

[0104] Calculate the ratio of the absolute value of the time derivative of the saccade modal component at the current moment to the preset saccade threshold, and map the ratio to the saccade compensation factor using a Gaussian decay function;

[0105] Multiply the noise suppression factor, the fidelity factor, and the saccade compensation factor at the current moment to obtain the confidence level at the current moment;

[0106] Construct a diagonal matrix using the confidence scores at each time point as diagonal elements, and use the diagonal matrix as the weighting matrix.

[0107] In some embodiments, the step of using a nullable neural network with piecewise non-monotonic activation functions to iteratively solve the time-varying error function to obtain a smooth trajectory includes:

[0108] A neural dynamics equation is constructed based on the principle of null neural networks, and the time-varying error function and time derivative are substituted into the neural dynamics equation.

[0109] Set a convergence threshold and a segmented interval for the activation function. During the iterative solution process, calculate the time-varying error function value at the current moment. Adjust the mapping parameters of the segmented non-monotonic activation function according to the segmented interval where the time-varying error function value is located.

[0110] The time-varying error function value is input into the piecewise non-monotonic activation function after parameter adjustment, and the state variables of the null neural network are updated.

[0111] The process continues iteratively until the time-varying error function value is less than the convergence threshold. The output state variable of the nullified neural network at this point is then used as the smoothed eye-tracking position coordinates, and the smoothed trajectory is formed based on the time series.

[0112] In some embodiments, calculating the weighted deviation between the fixation point and the corresponding position of the smoothed trajectory in the original eye movement trajectory data, and obtaining the corrected fixation point sequence after compensation, includes:

[0113] Obtain the original gaze point coordinates at the current moment in the original eye movement trajectory data, and the smooth point coordinates at the current moment in the smooth trajectory;

[0114] Calculate the vector difference between the original gaze point coordinates and the smoothed point coordinates to obtain the position deviation vector;

[0115] Calculate the ratio of the magnitude of the position deviation vector to the preset reference attenuation radius, calculate the attenuation value of the ratio using the exponential attenuation function, subtract the attenuation value from 1 to obtain the deviation weight, and multiply the position deviation vector by the deviation weight to obtain the weighted deviation.

[0116] Subtract the weighted deviation from the original fixation point coordinates to obtain the corrected fixation point coordinates at the current time. Sort the corrected fixation point coordinates at each time according to time to form the corrected fixation point sequence.

[0117] In some embodiments, calculating the deviation of the corrected fixation point sequence from the target position in the preset visual stimulus pattern, converting the deviation into a calculated visual angle error in conjunction with the set observation distance, and determining the visual acuity test result by combining the calculated visual angle error with the set optotype size includes:

[0118] Extract the true coordinate sequence of the preset target point at each time moment in the preset visual stimulus pattern;

[0119] Calculate the Euclidean distance between each fixation point in the corrected fixation point sequence and the true coordinates at the corresponding time, and use it as the instantaneous deviation error at each time.

[0120] The average deviation error is calculated by discretely summing the instantaneous deviation errors at all times and dividing by the total number of sampling points.

[0121] The average deviation error is converted into a calculated visual angle error by combining the set observation distance. The calculated visual angle error is compared with the preset stable fixation angle threshold. Based on the set target size and set observation distance corresponding to the preset visual stimulation mode when the stable fixation condition is reached, the corresponding visual acuity test level is determined and output as the visual acuity test result.

[0122] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art, upon reading this specification, may readily identify some of the devices as separate embodiments. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A vision detection method based on eye-tracking trajectory analysis, characterized in that, Includes the following steps: Acquire raw eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance; perform variational mode decomposition on the raw eye movement trajectory data to obtain multiple intrinsic mode functions, and identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral and energy characteristics. A noise suppression factor is determined based on the residual modal components, a fidelity factor is determined based on the reference modal components, and a saccade compensation factor is determined based on the saccade modal components. The product of the noise suppression factor, the fidelity factor, and the saccade compensation factor is used as the confidence level to construct a weighted matrix. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the goal of minimizing the trajectory fitting error modulated by the weighting matrix; a nulled neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory; The weighted deviation between the fixation point and the corresponding position of the smooth trajectory in the original eye movement trajectory data is calculated, and a corrected fixation point sequence is obtained after compensation. The degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern is calculated, and the degree of deviation is converted into a calculated visual angle error in combination with the set observation distance. The visual acuity test result is determined by combining the calculated visual angle error and the set optotype size.

2. The method according to claim 1, characterized in that, The variational mode decomposition operation on the original eye-tracking trajectory data to obtain multiple intrinsic mode functions includes: Set the penalty factor for variational mode decomposition and the number of intrinsic mode functions to be extracted; The decomposition process is constructed as a constrained variational extremum solution process, and the alternating direction multiplier method is used to update the center frequency and signal bandwidth of each decomposition state in the frequency domain one by one. Based on the updated center frequency, time-domain sequence reconstruction is performed in each frequency band, and isolated sequence data is stripped away to output the intrinsic mode functions corresponding to multiple channels without cross-aliasing characteristics.

3. The method according to claim 1 or 2, characterized in that, The process of identifying the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral and energy characteristics includes: Calculate the signal dominant frequency of each intrinsic mode function and the proportion of each intrinsic mode function in the overall energy distribution in a single extraction segment; The intrinsic mode functions with a dominant frequency of less than 2 Hz and an energy percentage of more than 50% are used as the reference mode components; The intrinsic mode function containing extreme velocity abrupt changes and exhibiting a broadband spectrum distribution is used as the scanned mode component; The intrinsic mode functions with a dominant frequency greater than 15Hz and frequency energy divergence exceeding a preset threshold are used as the residual mode components.

4. The method according to claim 1 or 2, characterized in that, The process of determining a noise suppression factor based on the residual mode components, a fidelity factor based on the reference mode components, and a saccade compensation factor based on the saccade mode components, and constructing a weighted matrix using the product of the noise suppression factor, the fidelity factor, and the saccade compensation factor as the confidence level, includes: The ratio of the absolute value of the residual modal component to the absolute value of the original eye-tracking trajectory data is used as the base ratio. The noise suppression factor is obtained by exponentiation with the natural constant as the base and the negative number of the base ratio as the exponent. Calculate the ratio of the time derivative of the reference mode component at the current moment to the preset reference velocity, perform a hyperbolic tangent transformation on the absolute value of the ratio, and subtract the transformation result from 1 to obtain the fidelity factor; Calculate the ratio of the absolute value of the time derivative of the saccade modal component at the current moment to the preset saccade threshold, and map the ratio to the saccade compensation factor using a Gaussian decay function; The confidence level at the current moment is obtained by multiplying the noise suppression factor, the fidelity factor, and the saccade compensation factor at the current moment. Construct a diagonal matrix using the confidence scores at each time point as diagonal elements, and use the diagonal matrix as the weighting matrix.

5. The method according to claim 4, characterized in that, The step of using a nullable neural network with piecewise non-monotonic activation functions to iteratively solve the time-varying error function to obtain a smooth trajectory includes: A neural dynamics equation is constructed based on the principle of null neural networks, and the time-varying error function and time derivative are substituted into the neural dynamics equation. Set a convergence threshold and a segmented interval for the activation function. During the iterative solution process, calculate the time-varying error function value at the current moment. Adjust the mapping parameters of the segmented non-monotonic activation function according to the segmented interval where the time-varying error function value is located. The time-varying error function value is input into the piecewise non-monotonic activation function after parameter adjustment, and the state variables of the null neural network are updated. The process continues iteratively until the time-varying error function value is less than the convergence threshold. The output state variable of the nullified neural network at this point is then used as the smoothed eye-tracking position coordinates, and the smoothed trajectory is formed based on the time series.

6. The method according to claim 1, characterized in that, The step of calculating the weighted deviation between the fixation point and the corresponding position of the smoothed trajectory in the original eye movement trajectory data, and obtaining the corrected fixation point sequence after compensation, includes: Obtain the original gaze point coordinates at the current moment in the original eye movement trajectory data, and the smooth point coordinates at the current moment in the smooth trajectory; Calculate the vector difference between the original gaze point coordinates and the smoothed point coordinates to obtain the position deviation vector; Calculate the ratio of the magnitude of the position deviation vector to the preset reference attenuation radius, calculate the attenuation value of the ratio using the exponential attenuation function, subtract the attenuation value from 1 to obtain the deviation weight, and multiply the position deviation vector by the deviation weight to obtain the weighted deviation. Subtract the weighted deviation from the original fixation point coordinates to obtain the corrected fixation point coordinates at the current time. Sort the corrected fixation point coordinates at each time according to time to form the corrected fixation point sequence.

7. The method according to claim 1, characterized in that, The calculation of the deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern, the conversion of the deviation into a calculated visual angle error based on the set observation distance, and the determination of the visual acuity test result based on the calculated visual angle error and the set optotype size include: Extract the true coordinate sequence of the preset target point at each time moment in the preset visual stimulus pattern; Calculate the Euclidean distance between each fixation point in the corrected fixation point sequence and the true coordinates at the corresponding time, and use it as the instantaneous deviation error at each time. The average deviation error is calculated by discretely summing the instantaneous deviation errors at all times and dividing by the total number of sampling points. The average deviation error is converted into a calculated visual angle error by combining the set observation distance. The calculated visual angle error is compared with the preset stable fixation angle threshold. Based on the set target size and set observation distance corresponding to the preset visual stimulation mode when the stable fixation condition is reached, the corresponding visual acuity test level is determined and output as the visual acuity test result.

8. A vision detection system based on eye-tracking trajectory analysis, characterized in that, Includes the following modules: The recognition module is used to acquire the original eye movement trajectory data when viewing a preset visual stimulus pattern that includes a set target size and a set viewing distance; to perform variational mode decomposition on the original eye movement trajectory data to obtain multiple intrinsic mode functions, and to identify the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral characteristics and energy characteristics. The solution module is used to determine a noise suppression factor based on the residual mode components, a fidelity factor based on the reference mode components, and a saccade compensation factor based on the saccade mode components. The product of the noise suppression factor, the fidelity factor, and the saccade compensation factor is used as the confidence level to construct a weighted matrix. Using the original eye-tracking trajectory data as input data, a time-varying error function is constructed with the goal of minimizing the trajectory fitting error modulated by the weighting matrix; a nulled neural network with piecewise non-monotonic activation functions is used to iteratively solve the time-varying error function to obtain a smooth trajectory; The determination module is used to calculate the weighted deviation between the fixation point in the original eye movement trajectory data and the corresponding position of the smooth trajectory, and obtain a corrected fixation point sequence after compensation; calculate the degree of deviation between the corrected fixation point sequence and the target position in the preset visual stimulus pattern, convert the degree of deviation into a calculated visual angle error in combination with the set observation distance, and determine the visual acuity test result by combining the calculated visual angle error and the set optotype size.

9. The system according to claim 8, characterized in that, The variational mode decomposition operation on the original eye-tracking trajectory data to obtain multiple intrinsic mode functions includes: Set the penalty factor for variational mode decomposition and the number of intrinsic mode functions to be extracted; The decomposition process is constructed as a constrained variational extremum solution process, and the alternating direction multiplier method is used to update the center frequency and signal bandwidth of each decomposition state in the frequency domain one by one. Based on the updated center frequency, time-domain sequence reconstruction is performed in each frequency band, and isolated sequence data is stripped away to output the intrinsic mode functions corresponding to multiple channels without cross-aliasing characteristics.

10. The system according to claim 8, characterized in that, The process of identifying the reference mode component of smooth following motion, the saccade mode component of saccade motion, and the residual mode component of noise based on spectral and energy characteristics includes: Calculate the signal dominant frequency of each intrinsic mode function and the proportion of each intrinsic mode function in the overall energy distribution in a single extraction segment; The intrinsic mode functions with a dominant frequency of less than 2 Hz and an energy percentage of more than 50% are used as the reference mode components; The intrinsic mode function containing extreme velocity abrupt changes and exhibiting a broadband spectrum distribution is used as the scanned mode component; The intrinsic mode functions with a dominant frequency greater than 15Hz and frequency energy divergence exceeding a preset threshold are used as the residual mode components.

Images